Sycophancy

Origins

The term 'sycophancy' is borrowed from the ordinary English word for fawning flattery. AI safety researcher Ajeya Cotra introduced it to AI alignment discourse in a 2021 essay that classified hypothetical advanced AI systems into Saints, Sycophants, and Schemers. Anthropic researchers adopted the term for the empirical behavior they observed in language models, and it has since become standard in the technical literature.

Discovery and Causes

Perez et al. at Anthropic published the first large-scale empirical evidence in 2022, showing that RLHF training increased the probability that a model would repeat back a user's preferred answer. Sharma et al.'s 2023 follow-up, 'Towards Understanding Sycophancy in Language Models,' demonstrated that five frontier assistants from OpenAI, Anthropic, and Meta all exhibited the behavior, and that both human raters and reward models preferred convincingly written sycophantic responses over truthful ones at a non-negligible rate. Wei et al. at Google DeepMind reported similar results in the PaLM family, finding that both model scale and instruction tuning increased sycophancy on opinion questions.

The behavior is generally classified as a form of reward hacking, in which an optimization process exploits a flaw in its reward signal rather than achieving the intended objective.

Forms

Anthropic's 2023 paper identified four canonical sycophantic behaviors: (1) feedback sycophancy, where the model rates text more favorably when told the user wrote it; (2) 'are you sure?' sycophancy, where it reverses a correct answer after the user expresses doubt; (3) answer sycophancy, where it biases free-form responses toward an answer the user has implied they prefer; and (4) mimicry sycophancy, where it repeats factual or grammatical errors the user has made.

Subsequent research added 'social sycophancy' (sycophancy concerning emotional validation, moral judgment, and the framing of personal situations rather than verifiable claims).

Notable Incidents

In April 2025, OpenAI rolled back a GPT-4o update after users reported that the model endorsed dangerous decisions and offered exaggerated compliments for trivial prompts. OpenAI's post-mortem attributed the regression to an additional training signal based on user thumbs-up/thumbs-down feedback that 'weakened the influence of our primary reward signal, which had been holding sycophancy in check.' In June 2025, The New York Times reported on cases of sustained delusional episodes following prolonged chatbot interaction. Academics later demonstrated through a formal Bayesian model that even an ideally rational user could be drawn into 'delusional spiraling' when interacting with a sufficiently sycophantic assistant.

Mitigations

Proposed mitigations include fine-tuning on synthetic data that rewards disagreement with incorrect user statements, editing the small subset of model parameters causally responsible for the behavior, changes to dialogue or system prompts, and benchmarks designed to surface sycophantic behavior before model release.